[lecture notes in computer science] learning and collaboration technologies. designing and...

Fuzzy Student Modeling for Personalization

of e-Learning Courses

Carla Limongelli and Filippo Sciarrone

Roma Tre University - Engineering DepartmentVia della Vasca Navale, 79

00146 Rome{limongel,sciarro}@dia.uniroma3.it

Abstract. In the context of e-learning courses, personalization is amore and more studied issue, being its advantage in terms of time andmotivations widely proved. Course personalization basically means tounderstand student’s needs: to this aim several Artificial Intelligencemethodologies have been used to model students for tailoring e-learningcourses and to provide didactic strategies, such as planning, case basedreasoning, or fuzzy logic, just to cite some of them. Moreover, in or-der to disseminate personalised e-learning courses, the use of known andavailable Learning Management System is mandatory.

In this paper we propose a fine-grained student model, embedded intoan Adaptive Educational Hypermedia, LS Plan provided as plug-in forMoodle. In this way we satisfy the two most important requirements: afine-grained personalization and a large diffusion. In particular, the sub-stantial modification proposed in this contribution regards the method-ology to evaluate the knowledge of the single student which currently hasa low granularity level. The experiments showed that the new system hasimproved the evaluation mechanism by adding information that studentsand teachers can use to keep track of learning progress.

1 Introduction

Distance learning is a mode of teaching/learning more and more required, used ineducation and working contexts. Research in this field has dramatically increasedin different directions: human-computer interfaces, design of Learning Manage-ment Systems (LMSs), social context [14], students’ modeling [20], teachers’background and teaching styles [18].

In this context,the use of known and available LMSs is mandatory and essen-tial to create and to spread content, but a LMS that provides all these featuresis difficult to find.

Moreover, the personalization of the learning experience is closely relatedto the efficiency and effectiveness of the learning process itself: personalizedcontent is more easily assimilated and the learner is more motivated. Howeverstudent diagnosis is uncertain, and a possible approach to face this problemis a fine-grained student modeling, that several researchers (e.g. [24]) assessed

P. Zaphiris and A. Ioannou (Eds.): LCT 2014, Part I, LNCS 8523, pp. 292–301, 2014.c© Springer International Publishing Switzerland 2014

Fuzzy Student Modeling for Personalization of e-Learning Courses 293

as adequate to carry out the student’s assessment and pedagogical strategies.Fuzzy logic techniques have been used to improve the performance of intelligenteducational systems due to their ability to handle uncertain information, suchas students actions, and to provide human descriptions of knowledge and ofstudents cognitive abilities [11].

In this paper we focus on the definition of a fine grained student model, takinginspiration from the work reported in [20], that allows to create and customizecourses basing on student’s learning styles according to the Felder-Silvermanmodel. We integrate into this model Kosba’s studies ([11]) on the application offuzzy logic for the description of the cognitive level of the students with respectto certain topics. We propose a plugin for the most world-widespread LMS: Moo-dle. In particular, our aim is to modify and improve the existing student model,that is the set of acquired information from student learning style and needs,and from the interactions with the system during her own personalized learningmaterial fruition. The current student model provides a student adaptive com-ponent, and evaluates students’ needs according to their knowledge and learningpreferences. In this proposal, the substantial modification regards the knowl-edge representation of the student, which currently has a low granularity level.LS Plan basically uses two values to estimate the knowledge: acquired or notacquired. We introduce in the new student model a fine-grained Cognitive Statewith four different levels of knowledge acquisition. This improvement brings agreat advantage to the end user that has her personalized course carefully tai-lored on her personal knowledge and learning goals, avoiding waste of time andmotivations in studying topics that could reveals trivial or too much difficult ina ”one-size-fits-all” course, or in a course not carefully personalized.

In the following section we report about some meaningful related work, in Sec.3 we describe the fine-grained Student Model. In Sec. 4 we show an example ofapplication. Conclusions and future work are drawn in Sec. 5.

2 Related Work

Fuzzy approach is a widely applied technique in user modeling, in particularit is used to model different aspects of student’s characteristics such as thedegree of knowledge the student has or acquires in a given subject. It allowsnatural description of knowledge and inference in the form of imprecise concepts,operators, and rules.

For example in [24] is proposed BSS1 that is an ITS with a general fuzzy logicengine for supporting the intelligent features of the ITS. In [2], ABITS definesthe domain knowledge where Learning Objects (LO) and related metadata arethe basic elements of the system. Its student model is represented by Cognitivestate, student’s preferences, and curricula. Fuzzy logic is applied here to rep-resent the uncertainty of the student’s score that can be obtained by differentkind of assessments and with different degree of reliability, as we also consider.Following a similar approach, in [12] the author presents TADV that is a famousframework for web distance education characterized by the Advice Generator

294 C. Limongelli and F. Sciarrone

that communicates with the student, during her didactic path, about resultsgained over time. Its student model is represented by four main aspects that, onthe whole, contribute to the overall student evaluation. They are: general infor-mation, behaviour, preferences, knowledge. The last one is represented by fuzzylogic, where the knowledge of a student about a certain topic is modeled by thefollowing variables: Certainty Factor (CF), Measure of Belief (MB), Measure ofDisbelief (MD). The Certainty Factor is a variable that takes on values between−1 and 1, and is directly influenced by the other two variables MB and MD.MB and MD are variables extrapolated from the results of the assessment quizthat the student takes for each concept of a specific course. Both MD and MBrange between 0 and 1. The first increases with each correct answer given by thestudent to a specific quiz. The second one is increased if the student respondsincorrectly or skips the quiz. The result of the two values is combined givingthe final value of CF. Kosba defines three levels of learning: completely learned,learned and unlearned. The CF is compared with the intervals that define thelearning levels and stores the result obtained by the student for the specificconcept. Not only quizzes modify the degree of understanding assigned to eachstudent, but also the time variable that contributes to define the Belief Graph.On the other way TADV gives suggestion to students, but does not provide apersonalized didactic strategy.

The work presented in [10] proposes an educational hypermedia where for eachdomain concept the user model saves a corresponding value using a linguisticvariable concept knowledge, which takes three possible values which estimatesuser knowledge of that concept: unknown, known, and learned. More recently,[3] proposes an adaptive learning environment for computer programming basedon hybrid student model, which combines an overlay model with stereotype andfuzzy logic techniques. In a perspective of technology enhanced learning, thereis research work [5], [4], [6], [21] aiming to integrate more traditional individu-alized e-learning [23] and social-collaborative e-learning [22], [1] and [7]. Manypersonalized e-learning systems do not use fuzzy approach to model students.They rather insist to suitably represent student’s learning styles, such as [9] or[20],[8], and also teachers’ teaching styles ([13], [15], [17],[16]).

We start from the work presented in [20], [19] and improve it with some aspectsof the student model that can be made more realistic by means fine-grainedmodeling. Being inspired by the work of Kosba, we integrate into LS Plan astudent model that will provide more realistic assessment of the students and thesubsequent re-planning of the learning that the student must follow. Unlike otherAdaptive Educational Systems like [10] and [3] that are ”problem oriented”,LS Plan provides a personalization engine that can be plugged in any educationalsystem being, therefore, problem independent.

3 The Student Model

In this section we present the Student Model (SM) and its relation with theadaptive system in which it is involved. Fig. 1 shows some relations among the


Fig. 1. Coarse-grained relations among components into Adaptive Systems. On thebases of a given SM, the Adaptation Engine configures a personalized course extractedfrom the Domain Knowledge. The course is presented to the student through a Hyper-media.

main components of an adaptive system. Before describing more in detail thecharacteristics of the new student model we will introduce some definitions aboutthe elements we are going to work with, then we show new mechanisms definedin the Adaptation Engine, which exploit the new model.

Definition 1. (Knowledge Item). A knowledge item KI is an atomic ele-ment of knowledge about a given topic. KI is a set: KI = {KIK ,KIA,KIE}where KI�, with � ∈ {K,A,E}, represents a cognitive level taken from Bloom’sTaxonomy: Knowledge, Application and Evaluation.

We have chosen only three out of the six levels of Bloom’s taxonomy: it is easy,but heavy, to provide the KI with all the six levels.

Definition 2. (Learning Style). A Learning Style LS is a 4-tuple: LS =〈D1, D2, D3, D4〉, withDi ∈ [−11,+11], i = 1, . . . , 4 where each Di is a Felder and Silverman LearningStyle Dimension, i.e., D1: active-reflective, D2: sensing-intuitive, D3: visual-verbal, D4: sequential-global.

We used the range [−11,+11] according to the Felder-Soloman ILS scale.

Definition 3. (Cognitive State). The Cognitive State CS is the set of allthe KI� possessed by the student with respect to the given topic: CS ⊆ DK.

Definition 4. (Student Model). The student model SM is a pair: SM =(CS,LS) where, CS is given in Definition 3 and LS is given in Definition 2.

This is the original Student Model definition presented in [20]. In the followingwe will not consider LS, because they will stay unchanged. Rather, we refineKIs, which contribute to CS, with fuzzy aspects taking inspiration from thework reported in [11].


Definition 5. (Fuzzy Knowledge Item). A fuzzy knowledge item KI is anatomic element of knowledge about a given topic. KI is a set: KI = {KIK ,KIA,KIE} where KI�, with � ∈ {K,A,E} is a real number. Each KI� ∈ [0, . . . , 1].The less the value, the less is the knowledge acquired for that item. KI is dividedin four evaluation classes: [0, 0.4) not sufficient, [0.4, 0.6) sufficient, [0.6, 0.8)good, [0.8, 1] excellent.

With this definitions we will represent always all the KI the student will dealwith, even if they are equal to 0 (see the outcome of the two SM in Fig. 3). Onthe other hand in the previous SM in CS there were only KI (fully) acquired.

The use of fuzzy logic is highlighted in the evaluation mechanism, in particularwhen a student is assigned a score to the end of a quiz. In order to describe thesemechanisms we have to introduce the Learning Node which is strictly related tothe quiz.

Definition 6. (Learning Node). A Learning Node LN is a 5-tuple: LN =〈LM,AK,RK,LS, T 〉 where

LM is the Learning Material, i.e., any instructional digital resource.AK Acquired Knowledge. It is a KI� that represents the knowledge that the

student acquires at a given level as specified in Definition 1, after havingpassed the assessment test related to the KI� of the node. If such a test isnot present in the node the AK is considered acquired anyway.

RK Required Knowledge. It is the set of KI� necessary for studying the materialof the node, i.e., the cognitive prerequisites required by the AK associatedto the node.

LS is given in Definition 2.T is a pair of reals T = (tmin, tmax) that represents the estimated time interval

for studying the material of the node, as prefixed by the teacher.

Let us consider a quiz related to a given LN . The mechanism for computingscores related to quizzes are normalized, for each quiz, between 0 and 1. Letus consider a LN having N questions related to it. The set of all the questionsQLN = {q1, . . . , qN} forms the quiz for that LN . Each qi has a weight w associ-ated to it, i.e. the score obtained in case of correct answer. The maximum scorea student can get in a quiz is 1.

N∑

k=1

w(qi) = 1

In this way, the score is directly related to the degree of knowledge acquired forthat LN . In case the teacher does not specify weights, the weight computationis given by

w(qi) =1

N


In particular, the new aspects of the student model will be the following: newrepresentation of the student’s knowledge; new representation of the score tests;new management of time variable associated to quizzes.

According to it, the learner’s knowledge will be classified within one of thesefour classes.

Timing are also considered in a different way: time computation is no morerelated to material fruition (that is difficult to estimate), but to the time thestudent needs to take the quiz at the end of each Learning Node. This is actu-ally more realistic evaluation than the time spent for studying the lesson itself.Obviously it is not possible to know if the fruition time (tf ) of a Learning Nodeis actually spent on studying or if it is affected by other factors. However, thethresholds tmin and tmax that we consider, allow to eliminate at least two stu-dent behaviors: the so-called ”coffee break” effect, when the fruition time tf isgreater than tmax, and a casual browsing of a given material, when tf is lessthan tmin.

This kind of knowledge representation provides also more flexibility for theteacher that can fix different thresholds for different LN , depending on theimportance that LN has in the context of a given course. For example if a LNexplains a very important concept that the student must acquire, the teachercan fix threshold to 0.7 and if less that LN will be proposed again to the student.

4 Experiments with Moodle

The new student model and the related management has been implemented intoLS Plan and Moodle 1.9. The experimentation has been carried out with thepurpose of verifying if the new student model is more refined than the previousone, highlighting the differences between the same two courses, the one runningwith improved LS Plan and the other one with the original plug-in. The experi-mental course is very small, but enough to show the difference with the previoussystem. Fig.2 shows the platform and the conceptual map of the experimentalcourse. In this experiment all the concepts are at the same level K.

We consider two courses with 5 Learning Nodes which represent the conceptsrelated to the course on neorealist cinema: intro C1, themes C2, movies C3,desica C4, rossellini C5. For the sake of readability in the table we show labelsassociated to the topics. 5 students attended the course. They first answeredto the initial questionnaire. For each concept the system selected a number ofquestions between 3 and 6. The overall initial questionnaire was composed by15 questions. The old system memorized the Cognitive State reported in thefollowing table:

The new system memorized also the level of knowledge that students got foreach concept in the KI Score section of the LS Plan Data Base.

We follow the student s2 whose initial situation is depicted in Fig. 2. Thefigure is a snapshot of the system after the initial questionnaire. Let us notethat only C2, C4 and C5 are proposed to the student, because he proved to knowalready concepts C1 and C3 in the initial questionnaire. s2 studies the first lessonproposed by the system, C2themes, obtaining a score of 0.7 as reported in Tab. 3.


Fig. 2. The experimental course. The interface of the old system and the new one isexactly the same, the difference is in the internal user knowledge representation and inthe adaptation mechanism at the initial stage.

Table 1. Cognitive state of the students after the initial questionnaire

Student Course Cognitive State CS Level

s1 id1 C1,C2,C5 K,K,K

s2 id1 C1,C3 K,K

s3 id1 C2,C4 K,K

s4 id1 C2 K

s5 id1 C2,C3,C5 K,K,K

Table 2. Cognitive state of the students after the initial questionnaire with the newstudent model. Bold numbers in the rightmost column represents the acquired concepts.

Student Course Cognitive State CS Level KI score (C1, . . . , C5)

s1 id1 C1,C2,C5 K,K,K 0.5,0.4, 0.2, 0, 0.5

s2 id1 C1,C3 K,K 0.5, 0.1, 0.5, 0.2, 0.2

s3 id1 C2,C4 K,K 0.2, 0.4, 0.1, 0.5, 0.2

s4 id1 C2 K 0.1, 0.5, 0.2, 0.2, 0.1

s5 id1 C2,C3,C5 K,K,K 0.1, 0.4, 0.4, 0.2, 0.5

The visible effect of the updating mechanism is an indication of the knowledgeacquired during the course (together with the learning preference that can betuned during the fruition of a course) as indicated by the system, like in Fig. 3.


Fig. 3. The outcome of the two advices. On the top there is an dialogue window thatdescribe user’s knowledge with the new model: all the KI are listed with their respectivedegree of acquisition. In the dialogue window below, the previous SM indicates onlythe list of acquired KI.

Table 3. KI score updating for concept C2 of student s2

Student Course Cognitive State CS Level KI score (C1, . . . , C5)

s2 id1 C1,C3,C5 K,K 0.5,0.7,0.5, 0.2, 0.2

5 Conclusions

In this work we adopted fuzzy representation for a student model used intothe Adaptive Educational System LS Plan. Starting form the student modelproposed in [20] and taking inspiration from the work presented in [12] we


introduced a fuzzy representation of the Cognitive State of the student modeland, consequently, all the related adaptation rules. These rules act mainly at thequiz level, in particular fuzzy CS modeling allows to give a more fine grainedevaluation about the knowledge acquired by the student and a consequent free-dom for the teacher to apply possible different didactic strategies.

In fact, with this new personalization process, not only teachers can estimatein a more precise way howmuch a student knows about a given topic, but also thestudent can have more appropriate suggestions about her/his way of learning.

The adaptation has been carried out with a plug-in for Moodle 1.9. We aregoing to adapt it to Moodle 2.5 in order to perform wider experiments.

References

1. Biancalana, C., Flamini, A., Gasparetti, F., Micarelli, A., Millevolte, S., Sansonetti,G.: Enhancing traditional local search recommendations with context-awareness.In: Konstan, J.A., Conejo, R., Marzo, J.L., Oliver, N. (eds.) UMAP 2011. LNCS,vol. 6787, pp. 335–340. Springer, Heidelberg (2011)

2. Capuano, N., Marsella, M., Salerno, S.: Abits: An agent based intelligent tutoringsystem for distance learning. In: Proceedings of the International Workshop inAdaptative and Intelligent Web-based Educational Systems, pp. 17–28 (2000)

3. Chrysafiadi, K., Virvou, M.: Dynamically personalized e-training in computer pro-gramming and the language c. IEEE Transactions on Education 56(4), 385–392(2013)

4. De Marsico, M., Sterbini, A., Temperini, M.: A strategy to join adaptive andreputation-based social-collaborative e-learning, through the zone of proximal de-velopment. International Journal of Distance Education Technologies 19(2), 105–121 (2012)

5. De Marsico, M., Sterbini, A., Temperini, M.: The definition of a tunneling strat-egy between adaptive learning and reputation-based group activities, pp. 498–500(2011)

6. De Marsico, M., Temperini, M.: Average effort and average mastery in the identi-fication of the zone of proximal development (2013)

7. Gasparetti, F., Micarelli, A., Sansonetti, G.: Exploiting web browsing activities foruser needs identification. In: Proceedings of CSCI 2014. IEEE Computer Society,Conference Publishing Services (March 2014)

8. Gentili, G., Micarelli, A., Sciarrone, F.: Infoweb: An adaptive information filteringsystem for the cultural heritage domain. Applied Artificial Intelligence 17(8-9),715–744 (2003)

9. Graf, S.: Kinshuk: Providing adaptive courses in learning management systemswith respect to learning styles. In: e-Learn 2007 (2007)

10. Kavcic, A.: Fuzzy user modeling for adaptation in educational hypermedia. IEEETransactions on Systems, Man, and Cybernetics, Part C: Applications and Re-views 34(4), 439–449 (2004)

11. Kosba, E.: Generating Computer-Based Advice in Web-Based Distance EducationEnvironments. PhD thesis, University of Leeds School of Computing (2004)

12. Kosba, E., Dimitrova, V., Boyle, R.: Using fuzzy techniques to model studentsin web-based learning environments. In: Palade, V., Howlett, R.J., Jain, L. (eds.)KES 2003. LNCS, vol. 2774, pp. 222–229. Springer, Heidelberg (2003)


13. Limongelli, C., Lombardi, M., Marani, A., Sciarrone, F.: A teacher model to speedup the process of building courses. In: Kurosu, M. (ed.) HCII/HCI 2013, Part II.LNCS, vol. 8005, pp. 434–443. Springer, Heidelberg (2013)

14. Limongelli, C., Lombardi, M., Marani, A., Sciarrone, F.: A teaching-style basedsocial network for didactic building and sharing. In: Lane, H.C., Yacef, K., Mostow,J., Pavlik, P. (eds.) AIED 2013. LNCS, vol. 7926, pp. 774–777. Springer, Heidelberg(2013)

15. Limongelli, C., Miola, A., Sciarrone, F., Temperini, M.: Supporting teachers toretrieve and select learning objects for personalized courses in the moodle-ls envi-ronment. In: Giovannella, C., Sampson, D.G., Aedo, I. (eds.) ICALT, pp. 518–520.IEEE (2012)

16. Limongelli, C., Mosiello, G., Panzieri, S., Sciarrone, F.: Virtual industrial training:Joining innovative interfaces with plant modeling. In: ITHET, pp. 1–6. IEEE (2012)

17. Limongelli, C., Sciarrone, F., Starace, P., Temperini, M.: An ontology-driven olapsystem to help teachers in the analysis of web learning object repositories. Infor-mation System Management 27(3), 198–206 (2010)

18. Limongelli, C., Sciarrone, F., Temperini, G., Vaste, M.: The lecomps5 frameworkfor personalized web-based learning: A teacher’s satisfaction perspective. Comput-ers in Human Beaviour 27(4) (2011)

19. Limongelli, C., Sciarrone, F., Vaste, G.: LS-plan: An effective combination of dy-namic courseware generation and learning styles in web-based education. In: Nejdl,W., Kay, J., Pu, P., Herder, E. (eds.) AH 2008. LNCS, vol. 5149, pp. 133–142.Springer, Heidelberg (2008)

20. Limongelli, C., Sciarrone, F., Vaste, G.: Personalized e-learning in moodle: Themoodle ls system. Journal of e-Learning and Knowledge Society 7(1), 49–58 (2011)

21. Sciarrone, F.: An extension of the q diversity metric for information processing inmultiple classifier systems: a field evaluation. International Journal of Wavelets,Multiresolution and Information Processing, IJWMIP 11(6) (2013)

22. Sterbini, A., Temperini, M.: Collaborative projects and self evaluation within asocial reputation-based exercise-sharing system, vol. 3, pp. 243–246 (2009)

23. Sterbini, A., Temperini, M.: Selection and sequencing constraints for personalizedcourses, pp. T2C1–T2C6 (2010); cited by (since 1996)

24. Warendorf, K., Tsao, S.J.: Application of fuzzy logic techniques in the bss1 tutoringsystem. J. Artif. Intell. Educ. 8(1), 113–146 (1997)

[lecture notes in computer science] learning and collaboration technologies. designing and...

Documents